Mapping AI at work

The artificial intelligence revolution has been remarkably selective in its reach across the workplace. Programmers are using it to generate and debug code, financial analysts to parse data, and administrators to automate routine written tasks. Yet across large stretches of the economy the technology remains largely a distant presence.

New research is beginning to put numbers on that divide. More than 80% of U.S. occupations now have at least one in five workers using generative AI, according to a September 2026 analysis from the Federal Reserve Bank of St. Louis. But intensive adoption is much rarer. In only 40% of occupations do a majority of workers use generative AI, while fewer than 3% of individual work tasks are performed with AI by a majority of workers.

This unevenness complicates predictions of an imminent transformation of employment. AI may be capable of performing large portions of many jobs, but capability is not the same as use. The first meaningful disruption may emerge where employers turn that potential into changes in hiring.

The workplace AI has actually reached

The popular image of AI adoption tends to divide the economy into jobs vulnerable to automation and jobs that are safe from it. Actual use produces a messier map.

The September 2026 St. Louis Fed analysis, based on nationally representative surveys of nearly 14,000 U.S. workers, found generative AI across an extraordinary range of occupations. Computer and information research scientists reported an adoption rate of 87.3%, with cybersecurity specialists close behind at 85.4%.

But even in the occupations most amenable to AI, there remains a considerable distance between what the technology can do and how extensively workers use it. The Anthropic March Economic Index examined that gap comparing the tasks AI could theoretically help perform with those for which workers were actually using Claude.

The discrepancy is striking. In computer and mathematical occupations, 94% of tasks are theoretically susceptible to AI assistance, compared with just 33% observed coverage on Claude. Computer programmers have the highest coverage at 75%, while data-entry keyers reach 67%. At the opposite extreme, 30% of U.S. workers are in occupations with zero measured Claude coverage, including cooks, mechanics, lifeguards, bartenders, and dishwashers.

Adoption largely follows the boundary of the computer screen. Work involving text, code and digital information remains far more accessible to generative AI than work rooted in physical environments.

What stops adoption

Some of the most revealing occupations are not those using AI most heavily, but those where its presence remains surprisingly small.

Medical secretaries and administrative assistants stand out for the gap between AI’s potential and its use. The September 2026 St. Louis Fed analysis puts adoption at just 16.8%, despite estimated exposure of 61%. Researchers point to privacy requirements, regulation and the cost of errors as possible barriers. Technical capability, in other words, does not guarantee adoption. Confidentiality, legal responsibility and institutional constraints can all determine whether AI is used in practice.

The same caution becomes more important when the lens moves outside the United States. The International Labour Organization's 2025 Generative AI and Jobs report estimates that one in four workers worldwide holds a job with some exposure to generative AI. Yet only 3.3% of global employment falls into its highest exposure category. Among women the share is 4.7%, compared with 2.4% among men, partly because women are more heavily represented in clerical occupations.

Occupational composition, wages and digital infrastructure vary enormously across economies, meaning the same AI capability can carry very different consequences in Beirut, Bangalore and Boston.

The youngest workers may be the first signal

The first signs of AI disruption may be appearing not among workers losing their jobs, but among younger workers trying to get one in the most exposed occupations.

Anthropic found no statistically significant increase in unemployment among workers in highly exposed U.S. occupations after the release of ChatGPT. Among workers ages 22 to 25, however, entry into the most exposed occupations declined by about half a percentage point per month relative to less-exposed jobs. Across the post-ChatGPT period, the job-finding rate for young workers entering exposed occupations was 14% lower than in 2022, although Anthropic cautions that the evidence remains tentative.

An August 2026 Stanford Digital Economy Lab study found a sharper divergence using payroll records from ADP, one of the largest payroll processors in the United States, covering millions of workers through June 2026. Employment among 22- to 25-year-olds in AI-exposed occupations stood 19% below where it would have been had it kept pace with employment among similarly aged workers in less-exposed occupations.

The mechanism may be as important as the number. Stanford cautions that the patterns do not establish causation, and the estimates weaken when educational attainment is taken into account. The emerging map of AI at work looks less like a uniform technological shock than a patchwork, advancing fastest through digital work. Its first effects on employment may emerge not in the disappearance of occupations, but in decisions about which workers companies hire in the first place.